def saved_model(testing_dataset,
model_id,
_run,
_log,
batching=False,
validation=False):
load_gdrive_file(model_id, 'zip')
def segmentation(image_path):
image = cv2.imread(image_path)
ZipFile(load_gdrive_file('12ONfO6WIS16xkfu6ucHEy4_5Tre0yxC5',
'zip')).extractall('/tmp/extracted_module')
tf.compat.v1.enable_resource_variables()
net = tf.saved_model.load('/tmp/extracted_module')
# batch processing
image = tf.expand_dims(image, 0)
out = net.signatures['serving_default'](tf.cast(image, tf.float32))
out = out['prediction'].numpy()[0]
# map colors
color_map = np.array([[128, 64, 128], [244, 35, 232], [70, 70, 70],
[102, 102, 156], [190, 153, 153], [153, 153, 153],
[250, 170, 30], [220, 220, 0], [107, 142, 35],
[152, 251, 152], [70, 130, 180], [220, 20, 60],
[255, 0, 0], [0, 0, 142], [0, 0, 70], [0, 60, 100],
[0, 80, 100], [0, 0, 230], [119, 11,
32]]).astype('int')
disp = color_map[out].astype('uint8')[..., ::-1] # convert to BGR for cv2
directory, filename = os.path.split(image_path)
cv2.imwrite(os.path.join(directory, f'{filename.split(".")[0]}_pred.png'),
disp)
def saved_model(testing_dataset,
model_id,
_run,
_log,
batching=False,
validation=False):
fsdata = FSData(**testing_dataset)
# Hacks because tf.data is shit and we need to translate the dict keys
def data_generator():
dataset = fsdata.validation_set if validation else fsdata.testset
for item in dataset:
data = fsdata._get_data(training_format=False, **item)
out = {}
for m in fsdata.modalities:
blob = crop_multiple(data[m])
if m == 'rgb':
m = 'image_left'
if 'mask' not in fsdata.modalities and m == 'labels':
m = 'mask'
out[m] = blob
yield out
data_types = {}
for key, item in fsdata.get_data_description()[0].items():
if key == 'rgb':
key = 'image_left'
if 'mask' not in fsdata.modalities and key == 'labels':
key = 'mask'
data_types[key] = item
data = tf.data.Dataset.from_generator(data_generator, data_types)
ZipFile(load_gdrive_file(model_id,
'zip')).extractall('/tmp/extracted_module')
tf.compat.v1.enable_resource_variables()
net = tf.saved_model.load('/tmp/extracted_module')
def eval_func(image):
if batching:
image = tf.expand_dims(image, 0)
out = net.signatures['serving_default'](tf.cast(image, tf.float32))
for key, val in out.items():
print(key, val.shape, flush=True)
return out['anomaly_score']
fs = bdlb.load(benchmark="fishyscapes", download_and_prepare=False)
_run.info['{}_anomaly'.format(model_id)] = fs.evaluate(eval_func, data)
def saved_model(testing_dataset, model_id, _run, _log, batching=True):
# testing_dataset is not used, but useful to keep config file consistent with other
# tests
data = tfds.load(
name='cityscapes',
split='validation',
data_dir='/cluster/work/riner/users/blumh/tensorflow_datasets')
label_lookup = tf.constant([
-1, -1, -1, -1, -1, -1, -1, 0, 1, -1, -1, 2, 3, 4, -1, -1, -1, 5, -1,
6, 7, 8, 9, 10, 11, 12, 13, 14, 15, -1, -1, 16, 17, 18
])
def label_lookup_map(batch):
batch['segmentation_label'] = tf.gather_nd(
label_lookup, tf.cast(batch['segmentation_label'], tf.int32))
return batch
data = data.map(label_lookup_map)
if batching:
data = data.batch(1)
ZipFile(load_gdrive_file(model_id,
'zip')).extractall('/tmp/extracted_module')
tf.compat.v1.enable_resource_variables()
net = tf.saved_model.load('/tmp/extracted_module')
m = tf.keras.metrics.MeanIoU(num_classes=19)
for batch in tqdm(data, ascii=True):
pred = net.signatures['serving_default'](tf.cast(
batch['image_left'], tf.float32))
labels = tf.reshape(batch['segmentation_label'], [-1])
weights = tf.where(labels == -1, 0, 1)
labels = tf.where(labels == -1, 0, labels)
m.update_state(labels,
tf.reshape(pred['prediction'], [-1]),
sample_weight=weights)
_run.info['mIoU'] = m.result().numpy()
def saved_model(testing_dataset, model_id, _run, _log, batching=False, validation=False):
data = tfds.load(name='cityscapes', split='validation',
data_dir='/cluster/work/riner/users/blumh/tensorflow_datasets')
if batching:
data = data.batch(1)
data = data.prefetch(500)
ZipFile(load_gdrive_file(model_id, 'zip')).extractall('/tmp/extracted_module')
tf.compat.v1.enable_resource_variables()
net = tf.saved_model.load('/tmp/extracted_module')
def eval_func(image):
out = net.signatures['serving_default'](tf.cast(image, tf.float32))
return out['anomaly_score'], out['prediction']
m = tf.keras.metrics.Mean()
for batch in tqdm(data, ascii=True):
start = time.time()
eval_func(batch['image_left'])
end = time.time()
m.update_state(end - start)
_run.info['{}_anomaly'.format(model_id)] = m.result().numpy()
def download(testing_dataset, batching, model_id):
load_gdrive_file(model_id, 'zip')
def saved_model(image_path,
name,
model_id,
scale='linear',
labels=None,
testing_dataset=None,
batching=False,
validation=None):
image = cv2.imread(image_path)
ZipFile(load_gdrive_file(model_id,
'zip')).extractall('/tmp/extracted_module')
tf.compat.v1.enable_resource_variables()
net = tf.saved_model.load('/tmp/extracted_module')
def eval_func(image):
if batching:
image = tf.expand_dims(image, 0)
out = net.signatures['serving_default'](tf.cast(image, tf.float32))
for key, val in out.items():
print(key, val.shape, flush=True)
return out['anomaly_score']
out = eval_func(image).numpy()
if batching:
out = out[0]
if scale == 'exp':
out = np.exp(out)
elif scale == 'log':
out = np.log(out)
min_val, max_val = out.min(), out.max()
disp = (out - min_val) / (max_val - min_val)
disp = 255 - (np.clip(disp, 0, 1) * 255).astype('uint8')
directory, filename = os.path.split(image_path)
if labels is None:
cv2.imwrite(
os.path.join(directory, f'{filename.split(".")[0]}_{name}.jpg'),
disp)
return
# since we have labels, check for the accuracy
def data_generator():
rgb = cv2.imread(image_path).astype('float32')
label = cv2.imread(labels, cv2.IMREAD_ANYDEPTH).astype('int32')
yield {'image_left': rgb, 'mask': label}
data = tf.data.Dataset.from_generator(data_generator, {
'image_left': tf.float32,
'mask': tf.int32
})
fs = bdlb.load(benchmark='fishyscapes', download_and_prepare=False)
metrics = fs.evaluate(eval_func, data)
print(metrics['AP'], flush=True)
cv2.imwrite(
os.path.join(
directory,
f'{filename.split(".")[0]}_{name}_AP{100 * metrics["AP"]:.2f}.jpg'
), disp)